Digital management method and system for medical documents

Abstract

The invention relates to the technical field of medical document management, in particular to a digital management method and system for medical documents. Obtaining medical document images of a target medical business system, processing the medical document images, and generating a standardized text stream of unified coding of each medical document image; constructing a feature matrix with spatial-semantic dual attributes in a feature region in each medical document through the standardized text stream of the unified coding of each medical document image; and extracting an entity relationship triple in the feature matrix, obtaining semantic fingerprint features of each medical document according to the entity relationship triple, and classifying each medical document according to the semantic fingerprint features of each medical document. According to the invention, efficient and accurate classification of the medical documents can be realized, and the intelligent and automatic level of medical document management is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical document management, and particularly to a digital management method and system for medical documents. Background Art

[0002] With the rapid development of information technology and the explosive growth of medical data volume, the traditional management method of paper medical documents has been difficult to meet the efficient and accurate data processing requirements of modern medical institutions. Medical documents include various forms such as medical records, prescription lists, imaging reports, and electronic inspection sheets. These documents not only carry the key medical information of patients but also involve complex structured and unstructured data. However, due to the diverse document formats, complex content, and strong semantic relevance, the traditional manual entry and classification methods are inefficient and prone to errors caused by human factors, making it difficult to achieve the standardized and intelligent management of medical information. In addition, the entity relationships (such as disease names, drug dosages, test indicators, etc.) and their spatial distribution characteristics (such as the layout and alignment of text blocks, etc.) contained in medical documents are crucial for the understanding and classification of documents. However, existing technologies often lack the ability to jointly analyze the spatial-semantic dual attributes, resulting in insufficient extraction and utilization of document information. Therefore, developing a digital management method that can efficiently process medical document images, extract structured information, and achieve intelligent classification has become an urgent need in medical informatization construction. Summary of the Invention

[0003] The present invention overcomes the deficiencies of the prior art and provides a digital management method and system for medical documents.

[0004] The technical solution adopted by the present invention to achieve the above object is as follows: In the first aspect of the present invention, a digital management method for medical documents is disclosed, including the following steps: Obtain medical document images of a target medical business system, process the medical document images to generate a standardized text stream with unified encoding for each medical document image; wherein, the medical document images include scanned medical records, prescription lists, imaging reports, and electronic inspection sheets; Construct a feature matrix with spatial-semantic dual attributes for feature regions in each medical document through the standardized text stream with unified encoding for each medical document image; Extract entity relationship triples from the feature matrix, obtain semantic fingerprint features of each medical document according to the entity relationship triples, and classify each medical document according to the semantic fingerprint features of each medical document.

[0005] Preferably, obtaining medical document images of a target medical business system, processing the medical document images to generate a standardized text stream with unified encoding for each medical document image is specifically as follows: Divide each medical document image into several candidate regions, classify and perform mask prediction on each candidate region, and output the class label and pixel-level mask of the corresponding structured region; Map the pixel-level mask back to the original image coordinate system to generate a partition mask with spatial coordinates for each medical document image, which is used to distinguish different structured regions; wherein, the structured regions include the table columns of the inspection form, the drug dosage area of the prescription, and the medical history description section of the medical record; Extract features from the partition mask with spatial coordinates of each medical document image through a Transformer model to obtain the page layout features and file header information of each medical document image; wherein, the page layout features include the text density of the page layout features, the table occupancy ratio, and the image distribution; the file header information includes the HL7 protocol identifier of the electronic inspection form and the DICOM tag of the imaging report; Establish a text stream index table including the source system, acquisition time, and document type for each medical document image based on the page layout features and file header information of each medical document image; Structurally process the text stream index table according to the clinical document architecture standard and perform standardized annotation using the SNOMED CT coding system; Use the feature data generated during the structural processing and standardized annotation as metadata tags, calculate the credibility weight values of each metadata tag through a dynamic weight algorithm, and finally generate a standardized text stream including the original image hash value, version number, and data traceability chain; wherein, the feature data includes the OCR recognition timestamp, correction operation log, and regional positioning coordinates.

[0006] Preferably, construct a feature matrix with dual spatial-semantic attributes for the feature regions in each medical document through the standardized text stream uniformly encoded for each medical document image, specifically: Obtain the relative coordinate parameter set of the feature regions in the medical document from the standardized text stream; wherein, the feature regions include the paragraph position of the scanned medical record, the signature area of the prescription form, the diagnosis conclusion box of the imaging report, and the inspection conclusion of the electronic inspection form; Extract the layout features of the feature regions according to the relative coordinate parameter set, construct a regional relationship representation based on the graph neural network according to the layout features of the feature regions, and obtain a three-dimensional spatial feature vector including the absolute position, relative layout, and visual saliency of the feature regions; Construct a dynamic semantic weight matrix, adjust the spatial position of the three-dimensional spatial feature vector according to the text position and entity type of each feature region in the corresponding medical document, and output a semantic feature vector with a hierarchical semantic structure; Taking each semantic feature vector as a node, establishing multiple edge connections according to each node to obtain a node adjacency graph, learning the spatial-semantic joint representation attention scores of each node in the node adjacency graph through a graph attention mechanism, and at the same time generating the attention weights of each node through spatial feature vectors; Corresponding the spatial-semantic joint representation attention scores of each node with the corresponding attention weights to obtain the semantic weight values of each node, and filling the semantic weight values of each node into the dynamic semantic weight matrix in a preset order; Using a non-linear mapping function to align the weight value scales of different modalities in the dynamic semantic weight matrix to generate a feature matrix with dual spatial-semantic attributes.

[0007] Preferably, using a non-linear mapping function to align the weight value scales of different modalities in the dynamic semantic weight matrix to generate a feature matrix with dual spatial-semantic attributes, specifically: Traverse each node in the dynamic semantic weight matrix, extract the original weight values of the spatial features and semantic features contained in the node respectively, and integrate the connection relationships between the nodes; According to a preset weight distribution model, determine the form of the non-linear mapping function and determine the function structure adapted to the data transformation of different modalities; Normalize the original weights of all nodes to generate an initial alignment matrix; after normalization, apply a non-linear mapping function to each modality respectively to adjust its numerical scale so that the weights of each modality are consistent on the same scale; According to the node adjacency graph, calculate the mutual influence between each node, and further correct the weight values output by the mapping function through a graph attention mechanism; Integrate the weight values of each node after non-linear mapping and graph attention adjustment into a new weight matrix in a preset order to obtain a feature matrix with dual spatial-semantic attributes; wherein, the numerical value of each position in the feature matrix reflects the dual attributes of the node in the spatial structure and semantic content.

[0008] Preferably, extract the entity relationship triples in the feature matrix, obtain the semantic fingerprint features of each medical document according to the entity relationship triples, and classify each medical document according to the semantic fingerprint features of each medical document, specifically: Extract the spatial coordinate distribution and semantic embedding vectors of the medical document feature region in the feature matrix, construct a bidirectional graph structure network based on the spatial coordinate distribution and semantic embedding vectors, establish an initial connection between each semantic unit node in the bidirectional graph structure network and its spatial neighborhood nodes, and dynamically adjust the connection weights of the bidirectional graph structure network according to the semantic similarity between the nodes; The cross - dimensional interaction calculation is performed on the bidirectional graph structure network using the multi - head attention mechanism to capture the explicit and implicit associations between medical entities and context - descriptive texts, and generate entity - relationship triples; among them, the medical entities include disease names, drug dosages, and test indicators. Analyze the entity types, location distributions, and relationship topological structures in the entity - relationship triples, and statistically analyze the high - frequency entity combination patterns in the corresponding medical documents according to the entity types, location distributions, and relationship topological structures in the entity - relationship triples, and extract the layout density, alignment method, and nested - level features of the feature regions of each medical document. Construct multi - dimensional feature vectors of the feature regions of each medical document according to the layout density, alignment method, and nested - level features of the feature regions of each medical document. The multi - dimensional feature vectors are used to represent the domain - specific vectors of each medical document; use a hierarchical dimensionality - reduction algorithm to perform spatial - semantic joint encoding on the multi - dimensional feature vectors of the feature regions of each medical document to generate the semantic fingerprint features of each medical document. Calculate the similarity between the semantic fingerprint features of each medical document and the preset semantic fingerprint features in several preset storage spaces. If the similarity between the semantic fingerprint feature of a certain medical document and the preset semantic fingerprint feature in a certain preset storage space is greater than the preset similarity threshold, then map the medical document into the preset storage space.

[0009] Preferably, the cross - dimensional interaction calculation is performed on the bidirectional graph structure network using the multi - head attention mechanism to capture the explicit and implicit associations between medical entities and context - descriptive texts, and generate entity - relationship triples. Specifically: Align the spatial coordinate distributions of each semantic unit node in the bidirectional graph structure network with the semantic embedding vectors to generate a composite feature tensor containing spatial position encoding and semantic vectors; among them, the spatial coordinate distributions include the center point coordinates of the text block and the boundary box size. Determine query vectors, key vectors, and value vectors for each node. The query vectors are used for entity - type recognition, including drug names and test items. The key vectors are used to fuse spatial proximity and semantic association strength, and the value vectors are used to carry node context features. Divide the composite feature tensor into multiple subspace heads, and each subspace head independently calculates the attention weights between nodes: calculate the relative position offset through the spatial coordinate difference to generate spatial relationship weights; calculate the semantic association weights based on the cosine similarity of the semantic embedding vectors to strengthen the potential connections of entities of the same category. Dynamically weight - fuse the spatial relationship weights and semantic association weights of each subspace head to generate a cross - dimensional attention distribution map. Weighted aggregation of the value vectors according to the multi-head attention weights to generate context-enhanced features for each node while preserving the original spatial layout information; Walk in the attention distribution map along the connection path with the highest attention weight to identify cross-region entity association chains and extract directional triples; Calculate the confidence score of the extracted triple pairs according to the context-enhanced features of the corresponding nodes through the multi-layer perceptron algorithm to obtain the confidence score of the extracted triples; Label the triples with a confidence score not greater than the preset threshold as invalid triples; label the triples with a confidence score greater than the preset threshold as valid triples, and combine all valid triples to form the final set of entity relationship triples.

[0010] A second aspect of the present invention discloses a digital management system for medical documents. The digital management system for medical documents includes a memory and a processor. A program for the digital management method of medical documents is stored in the memory. When the program for the digital management method of medical documents is executed by the processor, the steps of the digital management method of medical documents as described in any one of the above are implemented.

[0011] The present invention solves the technical defects in the background art and has the following beneficial effects: obtaining medical document images of a target medical business system, processing the medical document images to generate a standardized text stream with unified encoding for each medical document image; constructing a feature matrix with spatial-semantic dual attributes for the feature regions in each medical document through the standardized text stream with unified encoding for each medical document image; extracting entity relationship triples from the feature matrix, obtaining semantic fingerprint features of each medical document according to the entity relationship triples, and classifying each medical document according to the semantic fingerprint features of each medical document. The present invention can convert medical document images into a standardized text stream, construct a feature matrix with spatial-semantic dual attributes, extract entity relationship triples from it and generate semantic fingerprint features, thereby realizing efficient and accurate classification of medical documents and improving the intelligent and automated level of medical document management. Description of the Drawings

[0012] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0013] Figure 1 It is an overall method flowchart of a digital management method for medical documents; Figure 2 Part of the method process for a digital management method of medical documents; Figure 3 System block diagram of a digital management system for medical documents. Detailed implementation manners

[0014] In order to more clearly understand the above objects, features and advantages of the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners. It should be noted that, without conflict, the embodiments of the present application and the features in the embodiments may be combined with each other.

[0015] Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.

[0016] As Figure 1 shown, a first aspect of the present invention discloses a digital management method for medical documents, including the following steps: S102. Obtain medical document images of a target medical business system, process the medical document images, and generate a standardized text stream with unified encoding for each medical document image; wherein, the medical document images include scanned medical records, prescription drug lists, imaging reports, and electronic inspection lists; S104. Construct a feature matrix with spatial-semantic dual attributes for feature regions in each medical document through the standardized text stream with unified encoding for each medical document image; S106. Extract entity relationship triples from the feature matrix, obtain semantic fingerprint features of each medical document according to the entity relationship triples, and classify each medical document according to the semantic fingerprint features of each medical document.

[0017] The present invention can convert medical document images into a standardized text stream, construct a feature matrix with spatial-semantic dual attributes, extract entity relationship triples therefrom and generate semantic fingerprint features, thereby realizing efficient and accurate classification of medical documents and improving the intelligent and automated level of medical document management.

[0018] Preferably, obtaining medical document images of a target medical business system, processing the medical document images, and generating a standardized text stream with unified encoding for each medical document image, as Figure 2 shown, specifically: S202. Divide each medical document image into several candidate regions, perform classification and mask prediction on each candidate region, and output the category label and pixel-level mask of the corresponding structured region; S204. Map the pixel-level mask back to the original image coordinate system to generate a partition mask with spatial coordinates for each medical document image, which is used to distinguish different structured areas. The structured areas include the table columns of the inspection form, the drug dosage area of the prescription, and the medical history description section of the medical record. S206. Extract features from the partition mask with spatial coordinates of each medical document image through the Transformer model to obtain the page layout features and file header information of each medical document image. The page layout features include the text density of the page layout features, the table proportion, and the image distribution. The file header information includes the HL7 protocol identifier of the electronic inspection form and the DICOM tag of the imaging report. S208. Establish a text stream index table for each medical document image containing the source system, collection time, and document type based on the page layout features and file header information of each medical document image. S210. Structurally process the text stream index table according to the clinical document architecture standard and perform standardized annotation using the SNOMED CT coding system. S212. Use the feature data generated during the structural processing and standardized annotation as metadata tags, calculate the credibility weight values of each metadata tag through the dynamic weight algorithm, and finally generate a standardized text stream containing the original image hash value, version number, and data traceability chain. The feature data includes the OCR recognition timestamp, correction operation log, and regional positioning coordinates.

[0019] Among them, establishing a text stream index table for each medical document image containing the source system, collection time, and document type based on the page layout features and file header information of each medical document image is specifically as follows: comprehensively analyze the page layout features and file header information of each obtained medical document image. For the page layout features, carefully distinguish the specific feature manifestations such as text density, table proportion, and image distribution. For the file header information, accurately extract the key elements such as the HL7 protocol identifier of the electronic inspection form and the DICOM tag of the imaging report. Then, integrate the information extracted from these analyses according to a specific logic to clarify the specific category of the source system, the accurate record of the collection time, and the clear definition of the document type. Then, construct an ordered data structure based on the integrated information, fill in the key items such as the source system, collection time, and document type one by one, and form a text stream index table framework, so as to finally establish a complete and accurate text stream index table for each medical document image containing the source system, collection time, and document type.

[0020] Among them, structuring the text stream index table according to the clinical document architecture standard and standardizing the annotation using the SNOMED CT coding system are specifically as follows: Structuring each field in the text stream index table according to the clinical document architecture standard, clarifying the type and format requirements of each part of the information, and ensuring the standardization and consistency of the data structure. Then, using the SNOMED CT coding system to standardize the annotation of medical terms in the text stream index table, converting each entity (such as disease name, drug dosage, test index, etc.) into the corresponding SNOMED CT code to ensure the consistency and comparability of terms. Finally, reorganize the standardized data into a format that conforms to the clinical document architecture standard to form the final structured and standardized text stream index table.

[0021] It should be noted that, first of all, by dividing the medical document image into candidate regions and performing classification and mask prediction, different structured regions can be accurately identified, such as the table columns of the test form, the drug dosage area of the prescription, and the medical history description section of the medical record, etc., realizing the fine partitioning of the document image. Then, use the Transformer model (such as XLNet, ALBERT) to extract the features of the partition mask with spatial coordinates, and obtain the page layout features and file header information, covering key contents such as text density, table ratio, image distribution, and specific protocol identifiers and tags. Then, establish a text stream index table based on these features and information, and perform structured processing and standardized annotation. At the same time, calculate the credibility weight value of the metadata label with a dynamic weight algorithm, so as to generate a standardized text stream containing rich information (such as the original image hash value, version number, data traceability chain, etc.). Generally speaking, this series of operations realizes the efficient processing of medical document images, makes the medical document images have the characteristics of unified coding and standardization, provides a reliable and standardized basis for the subsequent classification of medical documents, and helps to improve the intelligent level and data quality of medical document management.

[0022] Preferably, construct a feature matrix with spatial-semantic dual attributes in the feature regions of each medical document through the standardized text stream with unified coding of each medical document image, specifically as follows: Obtain the relative coordinate parameter set of the feature regions in the medical document from the standardized text stream; among them, the feature regions include the paragraph position of the scanned medical record, the signature area of the prescription form, the diagnosis conclusion box of the imaging report, and the test conclusion of the electronic test form; Extract the layout features of the feature regions according to the relative coordinate parameter set, construct a regional relationship representation based on the graph neural network according to the layout features of the feature regions, and obtain a three-dimensional spatial feature vector of the feature regions including absolute position, relative layout, and visual saliency; Construct a dynamic semantic weight matrix, adjust the spatial position of the three-dimensional spatial feature vector according to the text position and entity type of each feature region in the corresponding medical document, and output a semantic feature vector with a hierarchical semantic structure; Taking each semantic feature vector as a node, establish multiple edge connections according to each node to obtain a node adjacency graph, and learn the spatial-semantic joint representation attention scores of each node in the node adjacency graph through a graph attention mechanism. At the same time, generate the attention weights of each node through the spatial feature vector; Correspond the spatial-semantic joint representation attention scores of each node with the corresponding attention weights to obtain the semantic weight values of each node, and fill the semantic weight values of each node into the dynamic semantic weight matrix in a preset order; Use a non-linear mapping function to align the weight value scales of different modalities in the dynamic semantic weight matrix, and generate a feature matrix with dual spatial-semantic attributes.

[0023] It should be noted that, first, obtain the relative coordinate parameter set of the feature region, which covers key regions such as the paragraph position of the scanned medical record. Then, based on this, extract the layout features and construct the regional relationship representation to obtain a three-dimensional spatial feature vector containing information such as the absolute position. Then construct a dynamic semantic weight matrix, adjust the position of the three-dimensional spatial feature vector according to the text position and entity type to obtain a semantic feature vector. Then establish multiple edge connections with the semantic feature vector as the node, and learn the spatial-semantic joint representation attention scores and generate attention weights of each node through the graph attention mechanism. Combine the two to obtain the node semantic weight value and fill it into the dynamic semantic weight matrix. Finally, use a non-linear mapping function to align the weight value scales, thereby generating a feature matrix that can simultaneously reflect the spatial features and semantic features. Generally speaking, the feature matrix constructed through this process can comprehensively reflect the spatial layout and semantic information of the feature region, provide a data basis and feature representation for subsequent in-depth understanding, classification, analysis, etc. of medical documents, and help improve the accuracy and intelligent level of medical document processing.

[0024] Preferably, use a non-linear mapping function to align the weight value scales of different modalities in the dynamic semantic weight matrix, and generate a feature matrix with dual spatial-semantic attributes, specifically: Traverse each node in the dynamic semantic weight matrix, extract the original weight values of the spatial features and semantic features contained in the node respectively, and integrate the connection relationships between the nodes; According to the preset weight distribution model, determine the form of the non-linear mapping function and determine the function structure suitable for data transformation of different modalities; Normalize the original weights of all nodes to generate an initial alignment matrix; after normalization, apply a non-linear mapping function to each modality respectively to adjust its numerical scale so that the weights of each modality are consistent on the same scale; According to the node adjacency graph, calculate the mutual influence between nodes, and further correct the weight values output by the mapping function through the graph attention mechanism; Integrate the weight values of each node after non-linear mapping and graph attention adjustment into a new weight matrix in a preset order to obtain a feature matrix with dual spatial-semantic attributes; wherein, the value at each position in the feature matrix reflects the dual attributes of the node in terms of spatial structure and semantic content.

[0025] It should be noted that by traversing the nodes of the dynamic semantic weight matrix, extracting the original weight values of spatial features and semantic features and integrating the connection relationships, this helps to comprehensively grasp the characteristics of the nodes. Determine the form of the non-linear mapping function and the structure suitable for data transformation of different modalities according to the weight distribution model to normalize the original weights, generate an initial alignment matrix, and then apply the non-linear mapping function to each modality to adjust the numerical scale to achieve the consistency of the weights of each modality on the same scale, so as to better fuse the information of different modalities. Then, calculate the mutual influence between nodes based on the node adjacency graph, and use the graph attention mechanism to further correct the weight values so that the weight values can better reflect the actual relationship between nodes. Finally, integrate the adjusted weight values into a new weight matrix, and the value at each position of the obtained feature matrix can reflect the dual attributes of the spatial structure and semantic content of the node at the same time. Through this process, the effective integration and adjustment of the weights of different modalities are realized, and the generated feature matrix can comprehensively and accurately reflect the spatial and semantic features of the nodes.

[0026] Preferably, extract the entity relation triples in the feature matrix, obtain the semantic fingerprint features of each medical document according to the entity relation triples, and classify each medical document according to the semantic fingerprint features of each medical document. Specifically: Extract the spatial coordinate distribution and semantic embedding vectors of the medical document feature region in the feature matrix, construct a bi-directional graph structure network based on the spatial coordinate distribution and semantic embedding vectors, establish an initial connection between each semantic unit node in the bi-directional graph structure network and its spatial neighborhood nodes, and dynamically adjust the connection weights of the bi-directional graph structure network according to the semantic similarity between nodes; Use the multi-head attention mechanism to perform cross-dimensional interaction calculations on the bi-directional graph structure network to capture the explicit and implicit associations between medical entities and context descriptive texts, and generate entity relation triples; wherein, the medical entities include disease names, drug dosages, and test indicators. Analyze the entity types, location distributions, and relationship topologies in the entity relationship triples, and count the high-frequency entity combination patterns in the corresponding medical documents according to the entity types, location distributions, and relationship topologies in the entity relationship triples, and extract the layout density, alignment method, and nested hierarchy features of the feature regions of each medical document; Construct a multi-dimensional feature vector of the feature regions of each medical document according to the layout density, alignment method, and nested hierarchy features of the feature regions of each medical document. The multi-dimensional feature vector is used to represent the domain-specific vector of each medical document; Use a hierarchical dimensionality reduction algorithm to perform spatial-semantic joint coding on the multi-dimensional feature vectors of the feature regions of each medical document to generate the semantic fingerprint features of each medical document; Calculate the similarity between the semantic fingerprint features of each medical document and the preset semantic fingerprint features in a number of preset storage spaces; If the similarity between the semantic fingerprint feature of a certain medical document and the preset semantic fingerprint feature in a certain preset storage space is greater than the preset similarity threshold, then map the medical document in the preset storage space.

[0027] It should be noted that relevant information is extracted from the feature matrix to construct a bidirectional graph structure network. By establishing initial connections and dynamically adjusting the connection weights, the spatial and semantic associations of the feature regions are reflected. Then, the multi-head attention mechanism is used for interactive calculation to capture the associations between medical entities (such as disease names, etc.) and the context text, thereby generating entity relationship triples. Then, the entity relationship triples are analyzed to obtain high-frequency entity combination patterns in the medical document and features such as the layout density of the feature regions. After that, multi-dimensional feature vectors are constructed based on these features, and then spatial-semantic joint coding is performed through a hierarchical dimensionality reduction algorithm to obtain semantic fingerprint features. Next, the similarity between the semantic fingerprint features of each medical document and the preset semantic fingerprint features is calculated. Finally, if the similarity is greater than the preset threshold, the medical document is mapped into the corresponding preset storage space, thus completing the automatic classification process of the medical document. Generally speaking, this method can extract key entity relationship triples from the feature matrix, and then obtain the unique semantic fingerprint features of each medical document, and based on this, accurately classify the medical documents; by constructing a bidirectional graph structure network, using attention mechanisms and other means, the spatial and semantic information of the medical document is fully mined, making the classification process more refined and accurate, and being able to effectively map medical documents with similar semantic features to specific storage spaces, which helps the efficient management and utilization of medical documents.

[0028] Preferably, use the multi-head attention mechanism to perform cross-dimensional interactive calculation on the bidirectional graph structure network to capture the explicit and implicit associations between medical entities and the context descriptive text, and generate entity relationship triples, specifically: Align the spatial coordinate distribution of each semantic unit node in the bidirectional graph structure network with the semantic embedding vector to generate a composite feature tensor containing spatial position encoding and semantic vectors; wherein, the spatial coordinate distribution includes the coordinates of the center point of the text block and the size of the bounding box. Determine a query vector, a key vector, and a value vector for each node. The query vector is used for entity type recognition, including drug names and test items. The key vector is used to fuse spatial proximity and semantic association strength. The value vector is used to carry the node context features. Divide the composite feature tensor into multiple subspace heads, and each subspace head independently calculates the attention weights between nodes: calculate the relative position offset through the spatial coordinate difference to generate the spatial relationship weight (such as the Gaussian attenuation coefficient of adjacent nodes); calculate the semantic association weight based on the cosine similarity of the semantic embedding vectors to strengthen the potential connection of entities of the same category (such as disease names and symptom descriptions). Dynamically weight and fuse the spatial relationship weight and the semantic association weight of each subspace head to generate a cross-dimensional attention distribution map; used to reveal the explicit dependencies between entity nodes (such as the vertical adjacent relationship of "drug name - usage dosage" in a prescription) and implicit associations (such as the semantic mapping of "index abbreviation - reference range" in a test sheet). Weightedly aggregate the value vectors according to the multi-head attention weights to generate the context-enhanced features of each node, while retaining the original spatial layout information. Walk in the attention distribution map along the connection path with the highest attention weight to identify the cross-region entity association chain (such as the horizontal logical link of "chief complaint - diagnosis - treatment plan" in a medical record), and extract the directional triples (subject → relation → object). Calculate the confidence score of the extracted triples pair according to the context-enhanced features of the corresponding nodes through the multi-layer perceptron algorithm to obtain the confidence score of the extracted triples. Calibrate the triples with a confidence score not greater than the preset threshold as invalid triples; calibrate the triples with a confidence score greater than the preset threshold as valid triples, and combine all valid triples to form the final entity relationship triple set.

[0029] Among them, through the multi-layer perceptron algorithm, the confidence score of the extracted triple pair is calculated based on the context-enhanced features of the corresponding nodes, and the confidence score of the extracted triple is obtained. Specifically: identify the extracted triple pair and obtain the context-enhanced features of the corresponding nodes; input the context-enhanced features of the corresponding nodes into the multi-layer perceptron algorithm, and based on the preset calculation logic and parameters, analyze and process the triple pair; by evaluating various factors such as the degree of association tightness, rationality, and fit with known patterns among the subject, relationship, and object in the triple, and combining the information provided by the context-enhanced features, gradually calculate a numerical value representing the confidence; this numerical value reflects the reliability and likelihood of the triple pair in the current context. After a series of iterative calculation processes, the confidence score of the extracted triple is finally determined, so as to accurately understand the credibility level of each triple.

[0030] It should be noted that the spatial coordinate distribution of the nodes in the bidirectional graph structure network (such as the center point coordinates of the text block and the bounding box size) is dimensionally aligned with the semantic embedding vector to form a composite feature tensor. Then, query vectors, key vectors, and value vectors for different functions are determined for each node. Then, the composite feature tensor is divided into multiple subspace heads, and the attention weights are calculated respectively. The spatial relationship weights are calculated through the spatial coordinate differences, the semantic association weights are calculated based on the semantic embedding vectors, and dynamic weighted fusion is performed to obtain the cross-dimensional attention distribution map to reveal various associations. Then, the value vectors are weighted and aggregated according to the multi-head attention weights to obtain the context-enhanced features of the nodes while retaining the spatial information. Then, the cross-region entity association chain is identified by walking along the path with high attention weights, and triples are extracted. The confidence scores of the triples are calculated through the multi-layer perceptron algorithm, and the invalid and valid triples are distinguished according to the scores. Finally, the valid triples are combined into the final entity relationship triple set. Generally speaking, the multi-head attention mechanism is used to realize the cross-dimensional interaction calculation of the bidirectional graph structure network, which can deeply capture the complex associations between medical entities and context texts, including explicit and implicit associations. The generated entity relationship triple set can accurately reflect the entity relationships and their confidence levels in the medical document, which helps to more deeply understand the semantics and structure of the medical document and improve the accuracy of medical document classification.

[0031] In this embodiment, the digital management method of the medical document further includes the following steps: Input the standardized text stream of the medical document into the pre-trained biomedical entity recognition network to scan the preset sensitive content in real time; among them, the preset sensitive content includes genetic information features (such as BRCA1 gene mutation markers) and mental illness keywords (such as ICD codes for bipolar disorder). Adopt an attention mechanism to weightedly evaluate the entity density and associated risk level of sensitive content (for example, mark the content involving the combination of genetic history and family relationship as the highest risk level), and generate a sensitivity score matrix of triples (entity type, risk weight, context relevance) with sensitivity scores; Obtain the sensitivity scores of the content of each medical document according to the sensitivity score matrix, and divide the content of each medical document into several levels according to the sensitivity scores of the content of each medical document, including first-level risk-level content, second-level risk content, and ordinary content; Automatically match the quantum key distribution protocol for first-level risk-level content (such as mental illness treatment records superimposed with genetic information), and enable the post-quantum encryption algorithm (NTRU lattice-based cryptography) for second-level risk content (single sensitive entity), and retain the national cryptographic standard encryption for ordinary content; When it is detected that a sensitive document is in the state of cross-institutional transmission or during the low-protection period at night, automatically raise the encryption intensity level and associate the device security authentication level of the document access terminal; When a document that needs quantum encryption is identified, connect to the quantum key distribution network in real time, and negotiate to generate a quantum true random key through the BB84 protocol; store the quantum key in physically isolated secure containers and blockchain smart contracts in slices, and only trigger key synthesis when both biometric authentication (such as doctor's fingerprint) and access policy verification (such as emergency rescue permission) are satisfied; dynamically adjust the key validity period according to the document sensitivity (genetic information is permanently valid, and ordinary diagnostic reports are decrypted within a limited time); Each time the document is changed, package the operator's digital identity, differential hash of the modified content, and timestamp fingerprint to generate an anti-tampering evidence block, and achieve second-level block confirmation through an improved DPoS-BFT hybrid consensus algorithm; when an unauthorized modification attempt of a sensitive document is detected, automatically trigger a cross-chain alarm contract, freeze the relevant account and start the data recovery process; Embed the visitor's identity characteristics (digital certificate number) and operation time (Beidou timekeeping code) into the document stream through quantum teleportation technology, and write the watermark information into the blockchain control chain at the same time; When a data leak occurs, accurately locate the leak link through the space-time matching of watermark decoding and blockchain logs (such as abnormal geographical location of a certain decryption operation); when the doctor's practice status changes or the patient withdraws the authorization, automatically trigger the smart contract to update the access policy, so that the new access requests for historical documents automatically upgrade the encryption level.

[0032] Generally speaking, this method realizes the full - life - cycle protection of highly sensitive medical data by deploying a dynamic encryption module based on diagnosis and treatment sensitivity, automatically enabling a quantum encryption channel for medical files containing preset content, and using blockchain technology to establish a document modification evidence chain. Each step is closely connected through a closed - loop control of real - time risk assessment, encryption policy matching, and blockchain evidence tracing, effectively solving the contradiction between the sharing and utilization of medical sensitive data and privacy protection. Among them, the medical files with the preset content include genetic information and mental diseases.

[0033] In this embodiment, the digital management method of the medical document further includes the method steps for tracing the contradictory information of the medical document and evaluating its credibility: Perform multi - layer comparison between the patient's latest document and the historical version stored in the blockchain. Among them, at the basic layer, the modified area is quickly located through the text hash value, and at the semantic layer, the medical entity relationship graph is used to identify key information changes (such as the addition or deletion of allergen names and dosage conflicts of medications). Construct a semantic coherence scoring model based on the bidirectional Transformer architecture. When it is detected that key entities (allergy history, past medical history) show logical contradictions in different versions (such as newly added penicillin allergy but the historical prescription contains amoxicillin records), automatically mark the conflict type (numerical conflict / logical contradiction / time paradox) and generate a differential vector matrix. According to the time range of the conflict mark, extract the encrypted meta - data packet of the relevant modification record (including the modifier's digital certificate, biometric signature hash, and Beidou - timed timestamp) from the medical evidence alliance chain. Through the light - node verification mechanism, synchronously verify the block - header information among multiple medical institution nodes to ensure the integrity and chronological authenticity of the traced data, and at the same time use zero - knowledge proof technology to protect non - essential privacy information. Input the encrypted biometric hash (such as doctor's fingerprint template, iris feature vector) stored in the blockchain into a quantum - safe container for decryption, and verify the authenticity of the signature source through a cross - institutional feature comparison model under the federated learning framework. Construct a multi - dimensional evaluation index system: including a basic dimension (modifier's permission level, digital certificate validity period), a behavior dimension (historical operation compliance rate, time - series pattern), and a technology dimension (biometric matching degree, block confirmation depth). Develop an adaptive weight assignment model: use the attention mechanism to dynamically adjust the weights of each dimension. For example, when it is detected that a doctor with high permissions modifies the document during non - working hours, automatically increase the biometric verification weight to 70%. Use the fuzzy logic algorithm to calculate the comprehensive credibility index, divide the risk level (credible / suspicious / high - risk), and generate a probabilistic score with a confidence interval. Convert the evaluation results into a tree - structured report containing triple evidence chains: (1) entity change chain (allergy history modification path map), (2) operator trajectory chain (heat map of historical operations of associated doctors), (3) timeline conflict matrix (comparison table of key time nodes); Present the propagation path of the modification record through a three - dimensional space - time sandbox, use color coding to mark the credibility levels of different versions, and support clicking to trace back to the original blockchain deposit block; recommend handling solutions based on the credibility index (green channel pass / yellow warning review / red alert freeze); Convert the misjudgment cases confirmed by manual review (situations where the system marks contradictions but are actually reasonable) into adversarial training samples, and optimize the threshold parameters of the semantic difference model through contrastive learning; when a certain type of contradiction pattern (such as the allergy history update specification for a specific department) appears repeatedly, automatically generate business rule patches and push them to each medical node; for high - risk evaluation results (such as credibility less than 30%), trigger the cross - institutional review process of associated documents in real - time and freeze the relevant modification permissions until multi - node consensus verification is completed.

[0034] Generally speaking, through four major technological breakthroughs of semantic difference positioning, blockchain space - time traceability, biometric cross - verification, and dynamic credibility modeling, this solution realizes the accurate traceability and intelligent evaluation of contradictory information in medical documents. Each step is closely connected through a closed - loop link of real - time conflict detection, distributed deposit verification, multi - dimensional evaluation, and feedback optimization, effectively solving the problems of information conflict verification and responsibility traceability in medical document version management, and providing technical support for the integrity protection of electronic medical records.

[0035] As Figure 3 shown, in the second aspect of the present invention, a digital management system 6 for medical documents is disclosed. The digital management system for medical documents includes a memory 41 and a processor 52. A program for the digital management method of medical documents is stored in the memory 41. When the program for the digital management method of medical documents is executed by the processor 52, the steps of any of the digital management methods of medical documents are implemented.

[0036] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only illustrative. For example, the division of the units is only a logical function division, and there can be other division methods in actual implementation. For example, multiple units or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed with each other can be through some interfaces. The indirect coupling or communication connection of devices or units can be electrical, mechanical, or other forms.

[0037] The units described above as separate components may or may not be physically separated, and the components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0038] In addition, in each embodiment of the present invention, all the functional units may be integrated in one processing unit, or each unit may be a separate unit alone, or two or more units may be integrated in one unit; the above-mentioned integrated units may be implemented in the form of hardware, or in the form of hardware plus software functional units.

[0039] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions. The foregoing program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps including the above method embodiments; and the foregoing storage medium includes: removable storage devices, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disks or optical disks and other various media that can store program codes.

[0040] Alternatively, if the above-mentioned integrated units of the present invention are implemented in the form of software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the embodiments of the present invention, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods of the various embodiments of the present invention. And the foregoing storage medium includes: removable storage devices, ROM, RAM, magnetic disks or optical disks and other various media that can store program codes.

[0041] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention.

Claims

1. A method for digital management of medical documents, characterized in that: The following steps are involved: Acquire medical document images of a target medical service system, process the medical document images, and generate a standardized text stream of uniformly coded medical document images; wherein the medical document images include scanned medical records, prescription drug orders, imaging reports, and electronic test orders; A feature matrix with dual spatial and semantic attributes of feature regions in each medical document is constructed through a standardized text stream uniformly encoded in each medical document image; Entity relationship triplets in the feature matrix are extracted, semantic fingerprint features of each medical document are obtained according to the entity relationship triplets, and each medical document is classified according to the semantic fingerprint features of each medical document.

2. A method for digital management of medical documents according to claim 1, characterized in that: Obtain medical document images of the target medical business system, process the medical document images, and generate standardized text streams of uniformly encoded medical document images, specifically: Each medical document image is divided into several candidate regions, each candidate region is classified and mask predicted, and the category label and pixel-level mask of the corresponding structured region are output; Mapping the pixel-level mask back to the original image coordinate system to generate a partition mask with spatial coordinates for each medical document image, which is used to distinguish different structured areas; wherein the structured areas include the table column of the test form, the drug dosage area of ​​the prescription, and the medical history description section of the medical record; The Transformer model is used to extract features from the partition masks with spatial coordinates of each medical document image to obtain the page layout features and file header information of each medical document image; wherein the page layout features include page layout feature text density, table proportion and image distribution; the file header information includes the HL7 protocol identifier of the electronic test form and the DICOM tag of the imaging report; Establishing a text stream index table for each medical document image including source system, acquisition time, and document type according to page layout features and file header information of each medical document image; Structuring the text stream index table according to the clinical document architecture standard and standardizing it using the SNOMED CT coding system; The feature data generated during the structured processing and standardized annotation process is used as metadata tags, and the credibility weight value of each metadata tag is calculated through a dynamic weight algorithm, and finally a standardized text stream containing the original image hash value, version number, and data traceability chain is generated; wherein the feature data includes OCR recognition timestamp, correction operation log, and regional positioning coordinates.

3. A method for digital management of medical documents according to claim 1, characterized in that: The feature matrix with dual spatial and semantic attributes of the feature regions in each medical document is constructed through the standardized text stream of the unified coding of each medical document image, specifically: Acquire a relative coordinate parameter set of a characteristic region in a medical document in the standardized text stream; wherein the characteristic region includes a paragraph position of a scanned medical record, a signature region of a prescription drug list, a diagnosis conclusion frame of an imaging report, and a test conclusion of an electronic test list; Extracting the layout features of the feature area according to the relative coordinate parameter set, constructing a regional relationship representation based on a graph neural network according to the layout features of the feature area, and obtaining a three-dimensional spatial feature vector of the feature area including absolute position, relative layout, and visual saliency; Constructing a dynamic semantic weight matrix, adjusting the spatial position of the three-dimensional spatial feature vector according to the text position and entity type of each feature region in the corresponding medical document, and outputting a semantic feature vector with a hierarchical semantic structure; Taking each semantic feature vector as a node, multiple edge connections are established according to each node to obtain a node adjacency graph, and the spatial-semantic joint representation attention score of each node in the node adjacency graph is learned through a graph attention mechanism, and the attention weight of each node is generated through the spatial feature vector; The spatial-semantic joint representation attention score of each node is matched with the corresponding attention weight to obtain the semantic weight value of each node, and the semantic weight value of each node is filled into the dynamic semantic weight matrix according to a preset order; A nonlinear mapping function is used to align the weight value scales of different modes in the dynamic semantic weight matrix to generate a feature matrix with spatial-semantic dual attributes.

4. A method for digital management of medical documents according to claim 3, characterized in that: A nonlinear mapping function is used to align the weight value scales of different modes in the dynamic semantic weight matrix to generate a feature matrix with dual spatial and semantic properties, specifically: Traversing each node in the dynamic semantic weight matrix, extracting original weight values ​​for the spatial features and semantic features contained in the nodes, and integrating the connection relationships between the nodes; According to the preset weight distribution model, the form of the nonlinear mapping function is determined, and the function structure adapted to the transformation of different modal data is determined; The original weights of all nodes are normalized to generate an initial alignment matrix. After normalization, a nonlinear mapping function is applied to each mode to adjust its numerical scale so that the weights of each mode are consistent on the same scale. According to the node adjacency graph, the mutual influence between the nodes is calculated, and the weight value output by the mapping function is further corrected through the graph attention mechanism; The weight values ​​of each node after nonlinear mapping and graph attention adjustment are integrated into a new weight matrix in a preset order to obtain a feature matrix with spatial-semantic dual attributes; wherein the value of each position in the feature matrix reflects the dual attributes of the node in spatial structure and semantic content.

5. A method for digital management of medical documents according to claim 1, characterized in that: Extract the entity relationship triples in the feature matrix, obtain the semantic fingerprint features of each medical document according to the entity relationship triples, and classify each medical document according to the semantic fingerprint features of each medical document, specifically: Extracting the spatial coordinate distribution and semantic embedding vector of the feature region of the medical document from the feature matrix, constructing a bidirectional graph structure network based on the spatial coordinate distribution and the semantic embedding vector, establishing an initial connection between each semantic unit node in the bidirectional graph structure network and its spatial neighboring nodes, and dynamically adjusting the connection weight of the bidirectional graph structure network according to the semantic similarity between the nodes; A multi-head attention mechanism is used to perform cross-dimensional interactive calculations on the bidirectional graph structure network to capture the explicit and implicit associations between medical entities and contextual descriptive texts, and generate entity relationship triples; wherein the medical entities include disease names, drug dosages, and test indicators; Analyze the entity type, location distribution and relationship topology in the entity relationship triples, count the high-frequency entity combination patterns in the corresponding medical documents according to the entity type, location distribution and relationship topology in the entity relationship triples, and extract the layout density, alignment and nested level features of the characteristic areas of each medical document; Constructing a multidimensional feature vector of the feature region of each medical document according to the layout density, alignment mode and nested hierarchical features of the feature region of each medical document, wherein the multidimensional feature vector is used to characterize the domain-specific vector of each medical document; performing spatial-semantic joint encoding on the multidimensional feature vector of the feature region of each medical document using a hierarchical dimensionality reduction algorithm to generate a semantic fingerprint feature of each medical document; Calculating the similarity between the semantic fingerprint feature of each medical document and the preset semantic fingerprint features in a plurality of preset storage spaces; If the similarity between the semantic fingerprint feature of a certain medical document and the preset semantic fingerprint feature in a certain preset storage space is greater than a preset similarity threshold, the medical document is mapped in the preset storage space.

6. A method for digital management of medical documents according to claim 5, characterized in that: The multi-head attention mechanism is used to perform cross-dimensional interactive calculations on the bidirectional graph structure network to capture the explicit and implicit associations between medical entities and contextual descriptive texts, and generate entity relationship triplets, specifically: Aligning the spatial coordinate distribution of each semantic unit node in the bidirectional graph structure network with the semantic embedding vector to generate a composite feature tensor including the spatial position encoding and the semantic vector; wherein the spatial coordinate distribution includes the coordinates of the center point of the text block and the size of the bounding box; Determine the query vector, key vector and value vector for each node, where the query vector is used to identify entity types, including drug names and test items, the key vector is used to integrate spatial proximity and semantic association strength, and the value vector is used to carry node context features; The composite feature tensor is divided into multiple subspace heads, and each subspace head independently calculates the attention weights between nodes: the relative position offset is calculated by the difference in spatial coordinates to generate the spatial relationship weight; the semantic association weight is calculated based on the cosine similarity of the semantic embedding vector to strengthen the potential connection between entities of the same category; The spatial relationship weight and semantic association weight of each subspace head are dynamically weighted and fused to generate a cross-dimensional attention distribution map; The value vectors are weighted aggregated according to the multi-head attention weights to generate context-enhanced features for each node while retaining the original spatial layout information; Walking along the connection path with the highest attention weight in the attention distribution map, identifying cross-region entity association chains, and extracting directional triplets; Calculate the confidence score of the extracted triplet pairs according to the context enhancement features of the corresponding nodes through a multi-layer sensing algorithm to obtain the confidence score of the extracted triplet; The triples whose confidence scores are not greater than the preset threshold are marked as invalid triples; the triples whose confidence scores are greater than the preset threshold are marked as valid triples, and all valid triples are combined to form the final entity relationship triple set.

7. A digital management system for medical documents, characterized in that: The digital management system for medical documents includes a memory and a processor. The memory stores a digital management method program for medical documents. When the digital management method program for medical documents is executed by the processor, the steps of the digital management method for medical documents as described in any one of claims 1 to 6 are implemented.

Images